Abstract: Accurate perception of surrounding rock conditions ahead of the tunnel face is difficult during TBM excavation because the face is shielded by the cutterhead and support system. To improve the rapid utilization of muck information and meet the real-time requirement of field applications, a lightweight ResNet-based method was developed for classifying TBM rock-chip states. Rock-chip images were collected from the Yangtze-to-Han River Water Diversion Project. According to the chainage, time records and surrounding-rock information, the apparent maximum block area  in each image was used as the primary indicator to divide the samples into three classes, namely normal excavation, parameter-adjustment condition and risk-warning condition. The images were preprocessed using OpenCV through brightness, saturation and contrast enhancement, together with edge-feature extraction. Based on standard residual blocks, nine lightweight ResNet candidates with different block combinations were constructed and compared on the original-image, enhanced-image and edge-image datasets. The results show that the standard ResNet34 suffered from severe overfitting under the small-sample condition, and the validation accuracy was lower than 40%. Image enhancement significantly improved the overall classification performance. The selected optimal model achieved an accuracy of 0.9688 on the training set and an overall accuracy of 0.87 on the test set, with F1-scores of 0.89, 0.85 and 0.87 for the three classes, respectively. The proposed method can effectively capture the differences in block size and particle-size distribution of rock chips and provides a practical basis for tunneling-parameter adjustment and surrounding-rock risk warning in TBM excavation.